Terminal multi-agent resource regulation method and system based on graph attention network

Abstract

The application discloses a terminal multi-agent resource regulation method and system based on a graph attention network, and the method comprises the following steps: a terminal process network model is constructed; a Markov game model is constructed, a dynamic regulation problem of multiple resource nodes is formalized into a multi-agent reinforcement learning framework, each agent corresponds to a resource node, resource regulation and configuration are performed according to local and upstream and downstream information, so that the average waiting time of passengers at each resource node is minimized and the balance of terminal resource configuration is achieved; a terminal multi-agent network structure perception mechanism is determined, the mechanism integrates adaptive feature selection and hierarchical attention perception mechanisms, and state feature representation considering local features and global structure is obtained; according to the obtained state features of the resource nodes, multi-agent intelligent decision generation is performed, and dynamic regulation decisions for the terminal resource nodes are obtained; the method improves the overall operation efficiency and stability of the terminal.

Description

Technical Field

[0001] This invention belongs to the field of resource collaborative regulation technology, and particularly relates to a multi-agent resource regulation method and system for airport terminals based on graph attention networks. Background Technology

[0002] With the rapid development of the civil aviation industry and the continuous growth of flight volume, the terminal building, as a key node for passenger travel, directly affects the overall service capacity and operational safety of the airport through its operational efficiency and resource allocation level. Terminals contain diverse resources, including check-in counters, security checkpoints, explosive ordnance detection gates, boarding gates, and rest areas. These different resource nodes have complex coupling relationships and dynamic passenger flow interactions. How to achieve efficient and orderly passenger flow under limited resource constraints has become an important research topic in the field of civil aviation operation management.

[0003] To address the above challenges, existing research has proposed various methods for terminal resource regulation and passenger flow management, which can be mainly divided into the following two categories: (1) static configuration methods based on rules or simulation; (2) scheduling methods based on prediction and optimization models. Static configuration methods based on rules or simulation: Early studies mostly adopted regulation strategies based on experience or rules, manually setting the number and time periods of open resources such as check-in counters and security checkpoints to achieve basic terminal operation guarantees; for example, by using queuing theory models or discrete event simulations to model the flow of passengers at each node and analyze the service efficiency under different configuration schemes, this type of method has a clear structure and is easy to implement, but because it relies on empirical rules, it is difficult to cope with Dynamic factors such as flight schedule fluctuations and random passenger behavior lead to low overall resource utilization. To improve the scientific rigor of resource allocation, some studies have introduced passenger flow prediction and optimization scheduling models. For example, these models predict passenger flow at each node based on regression analysis, time series analysis, or machine learning, and then optimize resource allocation using linear programming or integer programming. While these methods can achieve certain results in static or quasi-static scenarios, they require frequent calculations when facing real-time changes in passenger arrival characteristics, equipment failures, or unexpected events, resulting in insufficient real-time performance. Furthermore, these models often assume complete observability of the system state, making it difficult to handle the complex spatiotemporal coupling relationships within the terminal building.

[0004] In summary, although existing terminal resource management methods have made some progress in passenger queue management and resource allocation optimization, they still have significant shortcomings:

[0005] (1) Most methods fail to make full use of the topological information inside the terminal building; current control strategies often make decisions based on the state of local nodes, and the interaction between agents depends on simple neighborhood or global average information, which makes it difficult to accurately depict the hierarchical coupling and passenger flow relationship between different resource nodes, thus resulting in a lack of structural perception ability in system decision-making.

[0006] (2) Existing research focuses on optimizing the service efficiency of a single node, such as shortening the waiting time or improving the channel utilization rate, but ignores the synergistic effect and global balance between nodes, which can easily lead to local optima and overall performance degradation.

[0007] (3) Insufficient dynamic adaptability of the system; the passenger flow in the terminal is affected by multiple uncertain factors such as flight schedules, passenger behavior, and weather. The existing optimization model is slow to respond to sudden changes in passenger flow, resulting in a lag in resource allocation and affecting the overall operational efficiency.

[0008] In summary, existing terminal resource control technologies still suffer from problems such as insufficient structural perception, limited collaborative optimization capabilities, and weak dynamic adaptability, making it difficult to meet the needs of efficient and intelligent scheduling in complex terminal environments. Summary of the Invention

[0009] To address the problems existing in the background technology, this invention aims to propose a multi-agent resource regulation method and system for airport terminals based on graph attention networks, which addresses issues such as strong node heterogeneity, complex structural hierarchy, and insufficient global coordination in terminal resource regulation.

[0010] The present invention adopts the following technical solution:

[0011] A multi-agent resource regulation method for airport terminals based on graph attention networks, the method comprising the following steps:

[0012] Constructing a terminal building process network model includes: obtaining the terminal building process topology, constructing a directed graph structure of resource nodes, and performing passenger generation and behavior, process advancement, and queuing calculations.

[0013] A Markov game model is constructed to formalize the dynamic regulation problem of multiple resource nodes into a multi-agent reinforcement learning framework. The state space, observation space, action space, reward function and policy objective are determined. Under this multi-agent reinforcement learning framework, each agent corresponds to a resource node and performs resource regulation and configuration based on local and upstream and downstream information to minimize the average waiting time of passengers at each resource node and achieve the balance of terminal resource allocation.

[0014] A multi-agent network structure perception mechanism for the terminal building is determined. This mechanism integrates two mechanisms: adaptive feature selection and hierarchical attention perception. It can dynamically extract and aggregate key information based on the state of the resource nodes themselves and their hierarchy and adjacency relationships in the network, thereby obtaining a state feature representation that takes into account both local features and global structure.

[0015] Based on the obtained state characteristics of resource nodes, multi-agent intelligent decision generation is performed to obtain dynamic control decisions for terminal resource nodes.

[0016] Furthermore, adaptive feature selection mechanisms include:

[0017] Define the original feature vector based on the node type;

[0018] A feature selection attention mechanism is introduced, which automatically adjusts the importance of different features by calculating the correlation between node features and node type embeddings. This mechanism can be represented as:

[0019] ;

[0020] in, Represents resource nodes Feature selection attention weight vector; For resource nodes The original feature vector; For resource nodes Type embedding; Let be the learnable feature transformation matrix corresponding to the tanh activation function. This is the attention score parameter matrix before normalization using the softmax function; This represents a vector concatenation operation; Represents a nonlinear activation function; Represents the normalization function;

[0021] The differentiated feature representation of resource nodes is obtained through weighted operations:

[0022] ;

[0023] in, Represents resource nodes Differentiated weighted feature representation vector;

[0024] To achieve feature alignment between nodes of different types, a type-aware projection mechanism is further introduced to map differentiated features to a unified embedding space:

[0025] ;

[0026] in, Represents resource nodes The unified embedding representation vector obtained after type-aware projection; This represents a type-aware projection function; This represents a multi-layered sensing mechanism; This is the set of parameters for the MLP.

[0027] Furthermore, the hierarchical attention perception mechanism includes:

[0028] A unified embedding representation vector for each resource node Mapping to hidden representation:

[0029] ;

[0030] in, Represents resource nodes Hidden representation of nodes in the hidden space; is a learnable linear mapping parameter matrix;

[0031] To describe the hierarchical dependencies between nodes, a gating factor is introduced:

[0032] ;

[0033] in, Represents resource nodes With resource nodes The hierarchical perception gating coefficients between them; For the Sigmoid function; A vector of learnable parameters representing the gating factor; and Representing resource nodes respectively and resource nodes The hidden representation; and Representing resource nodes respectively With resource nodes The embedding vector of the corresponding level;

[0034] Based on this, we define attention weights:

[0035] ;

[0036] in, This indicates the neighboring resource nodes at the current time step. For resource nodes Hierarchical perception attention weights; This represents the learnable parameter vector in the attention mechanism; These represent the hidden representations of resource node k, respectively. Represents resource nodes The set of neighboring nodes in the terminal process network;

[0037] resource nodes The result of hierarchical sensing feature aggregation is as follows:

[0038] ;

[0039] in, Represents resource nodes Hierarchical perception feature representation after fusing multi-level structural information of neighboring nodes; This indicates a modified linear unit activation function;

[0040] Finally, the state feature representation of the resource nodes in the fused hierarchical structure is obtained. :

[0041] .

[0042] Furthermore, the reward function is:

[0043] ;

[0044] in, Representative resource node At time step The total reward value obtained, Represents local rewards, Represents collaborative rewards, Represents fair rewards;

[0045] in,

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] in, For resource nodes At time step Average waiting time at any given moment; This represents the total number of system resource nodes. This is the current resource node; This represents the global average latency. Indicates the current time step; This represents the number of discrete time steps in the statistics.

[0051] Furthermore, the state space is jointly composed of all the static and dynamic attributes of the resource nodes, and is represented as:

[0052] ;

[0053] in, This represents the set of resource nodes within the terminal building; Identifier for resource node type; For resource node level identification; Real-time queue length for resource nodes; This represents the average waiting time for the node. For the current resource node Number of open channels; For the current resource node The actual land area occupied; For the current resource node At time step The average floor space per passenger at any given time;

[0054] The observation space is used to form an observation vector for each resource node according to its own node type, and further integrates the state information from neighboring resource nodes to reflect the local operating state of the node at the current moment and its upstream and downstream relationships.

[0055] The action space is for adjusting the number of open channels of the current resource node, and is represented by a finite discrete set {-1,0,1} to reduce, maintain or increase one service channel;

[0056] The strategy aims to have each agent generate actions based on local observations, with the goal of maximizing the expected discounted reward.

[0057] Furthermore, based on the obtained state characteristics of the resource nodes, a multi-agent intelligent decision-making process is performed to obtain a method for dynamic control decisions for terminal resource nodes:

[0058] State information reception and fusion: At each decision time step, the agents of each functional node in the terminal receive state information from the environment;

[0059] Policy network decision generation: Input state features into the policy network agent, extract key decision features through deep neural structure and output the optimal control action; in the terminal resource control scenario, the agent outputs the action mainly to adjust the number of channels or counters that can be opened at the current node;

[0060] Action execution and environmental feedback: The agent applies the generated control actions to the terminal simulation model, and the operating status of the corresponding nodes changes accordingly; the environment calculates real-time feedback information based on the new operating status, including the reward values ​​of nodes and the whole system, and outputs the observation status of the next time step. This feedback reflects the impact of the current decision on passenger throughput efficiency and the overall system balance.

[0061] Experience trajectory recording and strategy optimization preparation: After completing a decision and interaction with the environment, the system will record the state, action, reward and next state of each agent, forming a complete experience trajectory sequence;

[0062] By repeating the above steps, the algorithm continuously updates and optimizes the strategy parameters in multiple rounds of iteration, eventually converging to the optimal control scheme. This enables the coordination and dynamic optimization of multiple resource nodes in the terminal, and obtains dynamic control decisions for the terminal resource nodes.

[0063] Furthermore, during process advancement and queue calculation,

[0064] The system advances the simulation in fixed time steps, dynamically updating passenger status and node actions; passengers Waiting time Defined as:

[0065] ;

[0066] Average wait time of resource nodes for:

[0067] ;

[0068] in, Indicates the passenger queue Passengers ahead, please gather. For passengers Service hours, For resource nodes The number of channels; Represents resource nodes The first One service channel This indicates the resource nodes within the current statistical time window. The A group of passengers gathered in the waiting area. Indicates channel The number of passengers included in the statistics.

[0069] A multi-agent resource regulation system for airport terminals based on graph attention networks includes:

[0070] The terminal process network model construction module is used to obtain the terminal process topology and construct a directed graph structure of resource nodes to perform passenger generation and behavior, process advancement and queuing calculation.

[0071] The Markov game model construction module is used to formalize the dynamic regulation problem of multiple resource nodes into a multi-agent reinforcement learning framework, which determines the state space, observation space, action space, reward function and policy objective. Under the multi-agent reinforcement learning framework, each agent corresponds to a resource node and performs resource regulation and configuration based on local and upstream and downstream information to minimize the average waiting time of passengers at each resource node and achieve a balance in terminal resource allocation.

[0072] The structure perception mechanism determination module is used to determine the structure perception mechanism of the terminal multi-agent network. The mechanism integrates two mechanisms: adaptive feature selection and hierarchical attention perception. It can dynamically extract and aggregate key information based on the state of the resource node itself and its hierarchy and adjacency relationship in the network, thereby obtaining a state feature representation that takes into account both local features and global structure.

[0073] The dynamic control decision acquisition module is used to generate multi-agent intelligent decisions based on the obtained state characteristics of resource nodes, and obtain dynamic control decisions for terminal resource nodes.

[0074] A non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-agent resource regulation method for airport terminals based on graph attention networks as described above.

[0075] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multi-agent resource regulation method for airport terminals based on graph attention networks as described above.

[0076] The beneficial technical effects of this invention are as follows:

[0077] (1) Existing terminal resource regulation research mostly relies on commercial simulation platforms to compare and analyze different resource configuration schemes through offline simulation, thereby obtaining empirical optimization conclusions. Such methods usually have problems such as low simulation efficiency, high cost of repeated experiments, and difficulty in supporting high-frequency interactive decision-making. Especially in the scenario of large-scale passenger flow and multi-node parallel regulation, it is difficult to meet the real-time or near-real-time optimization requirements. This invention constructs an autonomous and controllable terminal multi-agent simulation and decision-making framework, formalizes the resource regulation process into a multi-agent reinforcement learning interaction process, realizes high-frequency interaction between agents and the environment and online policy updates, significantly improves simulation operation efficiency and policy iteration speed, and provides a feasible technical path for dynamic resource optimization in complex scenarios.

[0078] (2) Existing resource regulation methods based on queuing theory, empirical rules or efficiency evaluation models usually rely on manually setting thresholds or static parameters, which makes it difficult to adjust resource allocation in real time according to changes in passenger flow and upstream and downstream pressures. They also have limited adaptability to sudden passenger flow, flight fluctuations and multi-node coupling effects. This invention enables each resource node agent to learn resource regulation strategies autonomously without explicitly modeling complex system dynamics through a multi-agent reinforcement learning mechanism. This achieves adaptive response to environmental changes, effectively reduces the subjectivity and limitations of manual rule design, and overcomes the dependence of rule-based or static model-based resource scheduling methods on human experience and prior assumptions.

[0079] (3) The resource nodes inside the terminal are highly coupled in space and process. Congestion in upstream nodes will have a significant pressure transmission effect on downstream nodes. Single node or local optimization strategies often lead to a decline in the overall performance of the system. This invention models the terminal resource regulation problem as a multi-agent Markov game and introduces a structure perception mechanism, so that each agent can not only consider its own state in the decision-making process, but also integrate the information of upstream and downstream nodes for coordinated regulation, effectively alleviating the problem of local congestion spreading to the whole process and improving the overall operating efficiency and stability of the terminal.

[0080] (4) Existing methods often focus on reducing the waiting time of a single node or local area, which can easily lead to uneven resource allocation or some nodes being under high load for a long time. This invention, by comprehensively considering the average passenger waiting time, node load and resource allocation balance in the reward function, guides the agent to achieve a dynamic trade-off between efficiency improvement and fair resource allocation under the premise of meeting the physical channel number constraint, thereby improving the overall robustness and service quality of terminal operation.

[0081] (5) The method proposed in this invention does not rely on a specific terminal structure or fixed rule configuration, can adapt to terminal scenarios of different scales and different process structures, and is easy to integrate with advanced algorithm modules such as graph learning and deep reinforcement learning. It has good versatility, scalability and engineering application potential. Attached Figure Description

[0082] Figure 1 This is the terminal operation process network model in Embodiment 1 of the present invention;

[0083] Figure 2 This is the overall framework for terminal resource regulation in Embodiment 1 of the present invention;

[0084] Figure 3(a) is a comparison of the average system waiting time of the method of the present invention provided in Example 2 and various typical multi-agent reinforcement learning algorithms under the same experimental conditions;

[0085] Figure 3(b) is a comparison of the maximum node waiting time of the method of the present invention provided in Example 2 and various typical multi-agent reinforcement learning algorithms under the same experimental conditions;

[0086] Figure 3(c) is a comparison of the fairness index of the method of the present invention provided in Example 2 and various typical multi-agent reinforcement learning algorithms under the same experimental conditions;

[0087] Figure 3(d) is a comparison of the volatility of the method of the present invention provided in Example 2 and various typical multi-agent reinforcement learning algorithms under the same experimental conditions;

[0088] Figure 4 Convergence performance analysis of the method of the present invention provided in Example 2 and various typical multi-agent reinforcement learning algorithms under the same experimental conditions;

[0089] Wherein, A2C represents the Independent Advantage Actor-Critic algorithm; IPPO represents the Independent Proximal Policy Optimization algorithm; MAPPO represents the Multi-Agent Proximal Policy Optimization algorithm; MAAC represents the Multi-Agent Attention Critic algorithm; and Ours represents the method of this invention. Detailed Implementation

[0090] The terminal multi-agent resource regulation method based on graph attention network provided by the present invention will be further described clearly and completely below with reference to the accompanying drawings:

[0091] Example 1

[0092] like Figure 1 As shown in the figure, the terminal multi-agent resource regulation method based on graph attention network provided in this embodiment includes the following steps:

[0093] S1. Constructing the terminal process network model; Constructing the terminal process network model includes: obtaining the terminal process topology, constructing a directed graph structure of resource nodes, and performing passenger generation and behavior, process advancement and queuing calculations.

[0094] Specifically,

[0095] S11. Obtain the terminal building process topology and construct a directed graph structure of resource nodes, thus building a terminal building operation process network model based on real flight data (e.g., Figure 1As shown in the figure, this model is based on an actual international airport and constructs a full-process network structure from check-in, explosive detection, security check, waiting to boarding. The nodes in the process network represent specific resource units, and the directed edges represent the flow paths of passengers between nodes. The service areas of the terminal are abstracted as resource nodes with adjustable channel numbers. A directed graph structure is constructed based on passenger flow relationships. Each node has local observation, decision-making and interaction capabilities, which are used to simulate the collaborative control process between intelligent agents. The system advances the passenger simulation according to a preset discrete time step (e.g., 30 minutes) and records key operational status indicators such as passenger arrival rate, queue length, and waiting time, thereby forming an environmental dynamic network model that can be used for policy learning.

[0096] S12. Passenger Generation and Behavior Calculation:

[0097] Historical airport operation data was obtained, using real operational data from the airport over the past two consecutive months (e.g., airport operation information from May to June 2025). This data covers approximately 25,000 departing flights and includes fields such as flight number, scheduled departure time, origin, destination, airline, and number of passengers. The obtained operational information data was preprocessed to ensure the accuracy of the input data and the reliability of the simulation environment. It should be noted that during the preprocessing of the acquired operational information data, the original flight data was filtered and cleaned to remove canceled flights and abnormal records. Missing passenger numbers were then filled in based on historical load factors. Passenger arrival processes were generated based on the flight schedule time distribution to reflect passenger flow fluctuations in different time periods. It is advisable to compare the generated passenger flow with official airport passenger flow statistics to verify the data's rationality.

[0098] Passenger arrival flow is dynamically generated based on flight schedules, and the passenger arrival process is modeled using a Poisson distribution.

[0099] (1)

[0100] in, Indicates the first The number of passengers arriving at the terminal within a given time frame. Indicates at time step Passenger arrival intensity;

[0101] Passengers follow the shortest queue selection strategy (SQS) when queuing between nodes:

[0102] (2)

[0103] in, Representing passengers at resource nodes The final target channel to choose from. For resource nodes The set of channels, For channel The real-time queue size;

[0104] S13, Process Execution and Queue Calculation:

[0105] The system advances the simulation in fixed time steps, dynamically updating passenger status and node actions; passengers Waiting time Defined as:

[0106] (3)

[0107] Average wait time of resource nodes for:

[0108] (4)

[0109] in, Indicates the passenger queue Passengers ahead, please gather. For passengers Service hours, For resource nodes The number of channels; Represents resource nodes The first One service channel This indicates the resource nodes within the current statistical time window. The A group of passengers gathered in the waiting area. Indicates channel The number of passengers included in the statistics;

[0110] S2. Construct a Markov game model to formalize the dynamic regulation problem of multiple resource nodes into a multi-agent reinforcement learning framework, providing a foundation for subsequent feature extraction and policy optimization. This invention models the terminal resource regulation problem as a Markov game. In this game, each agent corresponds to a resource node, and its task is to perform resource regulation and configuration based on local and upstream and downstream information under partially observable conditions, so as to minimize the average waiting time of passengers at each resource node and achieve a balance in terminal resource allocation, thereby realizing dynamic optimization of resources and global performance improvement.

[0111] The specific definitions are as follows:

[0112] state space It is composed of all the static and dynamic attributes of the resource node, that is:

[0113] (5)

[0114] in, This represents the set of resource nodes within the terminal building; This serves as a resource node type identifier, representing the resource node. The resource type is determined in advance before configuration and is read directly from the node configuration file during system initialization. This serves as a hierarchical identifier for resource nodes, representing resource nodes. The hierarchical position within the terminal's business process network should be predefined before configuration. The real-time queue length for resource nodes represents the time step. Time, resource node The total number of passengers queuing in all service channels is calculated by the system at each time step and updated in real time. The average waiting time of a node represents the time step. Inside, passengers are at resource nodes The average waiting time is calculated using formula (4); For the current resource node The number of open channels is counted and updated in real time at each time step by the system. For the current resource node The actual floor space is predefined before configuration; For the current resource node At time step The average floor space per passenger at any given time is calculated by the system at each time step using the following formula: Calculated and updated in real time;

[0115] Observation space Each intelligent agent (resource node) It will depend on its own node type This forms a corresponding observation vector, which is further fused with state information from adjacent resource nodes to reflect the local operating state of the node at the current moment and its upstream and downstream relationships.

[0116] Action space Adjust the number of open channels on the current resource node, using the finite discrete set {-1,0,1} to represent reducing, maintaining, or increasing one service channel;

[0117] reward function Taking into account average passenger waiting time, load conditions at upstream and downstream nodes, and the balance of terminal resource allocation; specifically,

[0118] This application constructs a collaborative reward mechanism that balances efficiency and load balancing: To achieve collaborative decision-making and dynamic resource optimization among resource nodes, a reward mechanism that integrates efficiency, collaboration, and fairness is designed.

[0119] (6)

[0120] in, Representative resource node At time step The overall reward value obtained measures the comprehensive effect of the current control measures on the overall terminal resource scheduling objectives. Represents local rewards, Represents collaborative rewards, Represents fair rewards;

[0121] This mechanism constrains agent behavior from three perspectives: single-node operating efficiency, global collaborative contribution, and system load balancing; among them, local rewards... (Efficiency items), Collaboration rewards (Global Items) Fairness Rewards The equilibrium terms are defined as follows:

[0122] (7)

[0123] in, For resource nodes At time step The average waiting time at any given moment is calculated using formula (4), and the shorter the time, the higher the reward.

[0124] (8)

[0125] (9)

[0126] in, This represents the total number of system resource nodes. This is the current resource node; The global average latency is the system's latency at each time step. Calculated according to the formula and updated in real time;

[0127] (10)

[0128] in, Indicates the current resource node; Indicates the current time step; This indicates the length of the time window used to calculate the fairness index, i.e., the number of discrete time steps in the statistics;

[0129] Strategy and Objective: Each agent is based on local observations Through strategy Generate an action with the goal of maximizing the expected discounted return:

[0130] (11)

[0131] in, This represents the number of discrete time steps in the statistics. As a discount factor, For the current resource node, This represents the total number of system resource nodes.

[0132] S3. Determine the multi-agent network structure perception mechanism for the terminal building to achieve adaptive perception and fusion of structural features in the multi-node resource network of the terminal building, so as to support efficient decision-making by agents in complex system environments. The multi-agent network structure perception mechanism integrates two mechanisms: adaptive feature selection and hierarchical attention perception. It can dynamically extract and aggregate key information based on the node's own state and its hierarchy and adjacency relationship in the network, thereby obtaining a representation that takes into account both local features and global structure. Specifically:

[0133] S31. Constructing an Adaptive Feature Selector for Differentiated Nodes: In view of the significant differences in functional attributes and operational characteristics of different types of resource nodes (such as check-in, security check, rest area, etc.) in the terminal process network, this invention proposes an Adaptive Feature Selector (AFS) to realize differentiated feature extraction and unified feature representation.

[0134] First, define the original feature vector according to the node type: For service-type nodes (such as check-in and security check), the main features include queue length, average waiting time, number of currently open channels and their upper limit, etc.; For capacity-type nodes (such as rest areas), the main features include area, number of people present and area per person, etc.

[0135] A feature selection attention mechanism is introduced, which automatically adjusts the importance of different features by calculating the correlation between node features and node type embeddings. This mechanism can be represented as:

[0136] (12)

[0137] in, Represents resource nodes The feature selection attention weight vector is used to measure the importance of each state feature of the resource node at the current decision moment. For resource nodes The original feature vector; For resource nodes Type embedding; Let be the learnable feature transformation matrix corresponding to the tanh activation function. This is the attention score parameter matrix before normalization using the softmax function; This represents a vector concatenation operation; This represents a non-linear activation function used to enhance the expressive power of feature maps; This represents the normalization function, used to map attention scores into a probability distribution, making the importance weights of different features comparable and summing to 1.

[0138] The differentiated feature representation of resource nodes is obtained through weighted operations:

[0139] (13)

[0140] in, Represents resource nodes Differentiated weighted feature representation vector;

[0141] To achieve feature alignment between nodes of different types, a type-aware projection mechanism is further introduced to map differentiated features to a unified embedding space:

[0142] (14)

[0143] in, Represents resource nodes The unified embedding representation vector obtained after type-aware projection; This represents a type-aware projection function used to determine the type based on resource nodes. The type performs a non-linear mapping on the input features to achieve alignment of the feature space of nodes across types; This represents a multi-layered sensing mechanism; For the MLP parameter set;

[0144] S32. Multi-level Information Aggregation in Terminals: This invention proposes a Hierarchical Gate Graph Attention Network (HG-GAT) to achieve feature aggregation and dependency modeling among multi-level nodes in a hierarchical heterogeneous network of a terminal; specifically:

[0145] A unified embedding representation vector for each resource node Mapping to hidden representation:

[0146] (15)

[0147] in, Represents resource nodes Hidden representation of nodes in the hidden space; is a learnable linear mapping parameter matrix;

[0148] To describe the hierarchical dependencies between nodes, a gating factor is introduced:

[0149] (16)

[0150] in, Represents resource nodes With resource nodes The hierarchical perception gating coefficients between nodes are used to dynamically control the strength of cross-layer information propagation between nodes; For the Sigmoid function; A vector of learnable parameters representing the gating factor; and Representing resource nodes respectively and resource nodes The hidden representation; and Representing resource nodes respectively With resource nodes Hierarchical embedding vector; gating coefficient Used for dynamically controlling the intensity of cross-layer information transmission;

[0151] Based on this, we define attention weights:

[0152] (17)

[0153] in, This indicates the neighboring resource nodes at the current time step. For resource nodes Hierarchical perception attention weights; This represents the learnable parameter vector in the attention mechanism; These represent the hidden representations of resource node k, respectively. Represents resource nodes The set of neighboring nodes in the terminal process network;

[0154] resource nodes The result of hierarchical sensing feature aggregation is as follows:

[0155] (18)

[0156] in, Represents resource nodes The hierarchical perception feature representation after fusing the multi-level structural information of neighboring nodes describes the characteristics of the resource node at the current moment after comprehensively considering the upstream and downstream processes and the status of related resource nodes. This represents the modified linear unit activation function, used to introduce nonlinear mapping capabilities and enhance the model's ability to express complex terminal operation states.

[0157] Finally, the state feature representation of the resource nodes in the fused hierarchical structure is obtained:

[0158] (19)

[0159] S4. Based on the state characteristics of the resource nodes obtained in step S3, multi-agent intelligent decision generation is performed. Based on the local state perceived by each node at the current moment, the information of neighboring nodes, and the global structure embedding, the system operation status is comprehensively judged, and corresponding control actions are output to obtain dynamic control decisions for terminal resource nodes. This process is geared towards terminal resource control scenarios, enabling intelligent agents of each resource node in the terminal to autonomously select the optimal control strategy in a complex and ever-changing environment, thereby achieving efficient collaboration of the overall system.

[0160] Specifically, the structure-aware multi-agent dynamic decision-making process for airport terminals is as follows:

[0161] S41. State Information Reception and Fusion: At each decision time step, agents at various functional nodes in the terminal (such as check-in islands, security checkpoints, waiting areas, etc.) receive state information from the environment; including:

[0162] a) Local operational characteristics of a node: such as the number of passengers, the number of people queuing, the average waiting time, and the number of available service channels at the current resource node;

[0163] b) Characteristics of upstream and downstream node associations: such as the flow input of upstream passengers to the current resource node, the load level of downstream nodes, etc.

[0164] c) Global structural embedding features: The graph attention representation obtained in step S3 reflects the overall operational status of the terminal and the dependencies between nodes; each node agent aggregates multi-source information through the feature fusion layer to obtain comprehensive state features that take into account its own state and the influence of upstream and downstream; this state feature representation not only reflects the current operating load of the node, but also reflects its structural position and potential bottleneck risks in the whole system.

[0165] S42. Policy Network Decision Generation: Input the state features into the policy network agent, extract key decision features through deep neural structure and output the optimal control action; in the terminal resource control scenario, the action output by the agent is mainly to adjust the number of channels or counters that can be opened at the current node;

[0166] S43. Action Execution and Environmental Feedback: The agent applies the generated control actions to the terminal simulation model, and the operating status of the corresponding nodes changes accordingly. The environment calculates real-time feedback information based on the new operating status, including the reward values ​​of the nodes and the global system, and outputs the observation status of the next time step. This feedback reflects the impact of the current decision on passenger throughput efficiency and the overall system balance.

[0167] S44. Experience Trajectory Recording and Policy Optimization Preparation: After completing a decision-making interaction with the environment, the system will record the state, actions, rewards, and next state of each agent, forming a complete experience trajectory sequence. These trajectories serve as input data for the subsequent agent optimization training phase, used to estimate policy advantages and perform parameter updates, thereby achieving continuous improvement and adaptive optimization of the agent's decision-making strategy.

[0168] Specifically, such as Figure 2 As shown, by using the constructed terminal simulation platform as the interactive environment, the intelligent agent sequentially completes environmental state perception, strategy decision-making, and control intervention in each simulation cycle, realizing a closed-loop optimization process of "simulation-perception-decision-intervention". The algorithm continuously updates and optimizes the strategy parameters in multiple cycles, and finally converges to the optimal control scheme, realizing the coordination and dynamic optimization of multiple resource nodes in the terminal.

[0169] To achieve efficient learning and stable convergence of multi-agent systems in complex terminal environments, this invention employs a proximal policy optimization (PPO) algorithm based on policy gradients during agent iterative optimization, cyclically updating the policies of each agent. This algorithm constrains the update magnitude between old and new policies, preventing training oscillations caused by excessively large update steps, thereby ensuring the stability and convergence of the policy optimization process. It is suitable for complex environments with high randomness and dynamic coupling characteristics, such as those found in terminal scenarios.

[0170] In the optimization process, the agent's time step is first calculated. Temporal Difference (TD) error :

[0171] (20)

[0172] in, Indicates an immediate reward. Representing state The value estimate, Discount factor;

[0173] Generalized Advantage Estimation (GAE) is introduced to improve the estimation accuracy of the advantage function. GAE is achieved by introducing a decay factor on the time difference error. Achieving a balance between bias and variance, its dominance function estimation The format is as follows:

[0174] (twenty one)

[0175] in, Indicates relative to the current time step The forward time offset steps are used to accumulate the time difference error over multiple future time steps.

[0176] The iterative optimization module optimizes based on the shear probability ratio. Construct a near-end policy objective function, and achieve stable updates by limiting its range of variation, thereby optimizing the objective function. as follows:

[0177] (twenty two)

[0178] in, For the new strategy and the old strategy in the state Select action The probability ratio, Indicates by parameters The current policy function is represented. This represents the old strategy function before the update; Indicates at time step The advantage function estimate obtained from environmental feedback is calculated by formula (21) to measure the superiority of the current action relative to the average strategy. The shear threshold; This represents the shear operator, when the input value is... Less than the lower bound The time value is When input value Greater than the upper bound The time value is ,when Located in the interval The internal time remains unchanged.

[0179] The objective function suppresses excessive policy update amplitude through shearing operations, thereby improving the learning efficiency and generalization ability of the agent in complex terminal environments while ensuring optimization stability.

[0180] Example 2

[0181] In this embodiment, to comprehensively reflect the operational efficiency, fairness, and regulatory stability of the framework proposed in this invention, the following four evaluation indicators are set:

[0182] a) System Average Wait (SAW) measures the average queuing time of global resource nodes, reflecting the overall operating efficiency of the system; its calculation formula is:

[0183] (twenty three)

[0184] b) Maximum Node Wait (MNW): This characterizes the peak waiting time of the most congested node in the system, reflecting the bottleneck characteristics of the system under high load scenarios; it is defined as:

[0185] (twenty four)

[0186] c) Fairness Index (FI), used to measure the degree of balance in service distribution among resource nodes, is defined using Jain's Fairness Index:

[0187] (25)

[0188] d) Wait Variance (WV), used to evaluate the stability of the system and the convergence characteristics of the strategy, is defined as:

[0189] (26)

[0190] To verify the performance advantages of this invention, the following typical algorithms were selected as a comparison baseline:

[0191] a) Independent Proximal Policy Optimization Algorithm (IPPO): Each agent is trained independently under completely decentralized conditions, and parameters are not shared. It is used to evaluate the autonomous decision-making ability and convergence performance under the constraint of no global information.

[0192] b) Multi-Agent Proximal Policy Optimization Algorithm (MAPPO): It adopts a framework of centralized training and decentralized execution, and uses shared global information to achieve centralized value estimation. It is used to verify the effect of centralized information fusion on improving policy stability and global coordination performance.

[0193] c) Advantage Actor-Critic (A2C): Based on the advantage function, the independent Actor-Critic structure has high update efficiency and structural simplicity. It is used to verify the adaptability and performance of traditional policy gradient methods in terminal resource regulation scenarios.

[0194] d) Multi-Agent Attention Critic (MAAC): This algorithm uses an attention mechanism to weighted aggregate information from neighboring agents, and is used to evaluate the role of explicit neighbor modeling in improving the performance of multi-agent interaction and cooperative control.

[0195] To verify the effectiveness and superiority of the present invention, based on the constructed multi-agent terminal simulation model, the control performance of the method of the present invention and a variety of typical multi-agent reinforcement learning algorithms (including the above-mentioned IPPO, MAPPO, A2C and MAAC) was compared and evaluated under the same experimental conditions; the experimental results are shown in Table 1 and Figures 3(a), 3(b), 3(c) and 3(d).

[0196] Overall results show that the method of this invention achieves the best performance in four core indicators: system average waiting time (SAW), maximum node waiting time (MNW), fairness index (FI), and volatility (WV). Specifically, the lowest system average waiting time and a significantly reduced maximum node waiting time indicate that the method effectively alleviates localized congestion during peak hours; the highest fairness index indicates more balanced resource allocation; and the lowest waiting time volatility reflects the stability and robustness of the system operation.

[0197] The performance improvement is mainly attributed to the Hierarchical Perceptual Graph Attention Network (HG-GAT) designed in this invention, which can efficiently model and aggregate information about the multi-level node relationships in the terminal's hierarchical structure, enabling the agent to consider both local state and neighborhood information when making decisions. Simultaneously, the collaborative reward mechanism introduces global load balancing constraints on top of traditional individual rewards, effectively promoting cooperation between nodes and thus achieving overall performance improvement.

[0198] Furthermore, the convergence analysis results show (see...) Figure 4 The method of this invention converges quickly during training, reaching a stable average return around the 80th training cycle, with minimal curve fluctuations. This indicates that the strategy optimization framework constructed in this invention possesses strong learning stability and generalization ability, maintaining stable resource control performance under varying passenger flow conditions across multiple scenarios.

[0199] Table 1 Comparison of four evaluation metrics for various algorithms

[0200]

[0201] Example 3

[0202] This embodiment provides a terminal multi-agent resource control system based on graph attention networks, including:

[0203] The terminal process network model construction module is used to obtain the terminal process topology and construct a directed graph structure of resource nodes to perform passenger generation and behavior, process advancement and queuing calculation.

[0204] The Markov game model construction module is used to formalize the dynamic regulation problem of multiple resource nodes into a multi-agent reinforcement learning framework, which determines the state space, observation space, action space, reward function and policy objective. Under the multi-agent reinforcement learning framework, each agent corresponds to a resource node and performs resource regulation and configuration based on local and upstream and downstream information to minimize the average waiting time of passengers at each resource node and achieve a balance in terminal resource allocation.

[0205] The structure perception mechanism determination module is used to determine the structure perception mechanism of the terminal multi-agent network. The mechanism integrates two mechanisms: adaptive feature selection and hierarchical attention perception. It can dynamically extract and aggregate key information based on the state of the resource node itself and its hierarchy and adjacency relationship in the network, thereby obtaining a state feature representation that takes into account both local features and global structure.

[0206] The dynamic control decision acquisition module is used to generate multi-agent intelligent decisions based on the obtained state characteristics of resource nodes, and obtain dynamic control decisions for terminal resource nodes.

[0207] A non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the multi-agent resource regulation method for airport terminals based on graph attention networks as described above.

[0208] Furthermore, the present invention adopts the following technical solution:

[0209] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the multi-agent resource regulation method for airport terminals based on graph attention networks as described above.

[0210] From the above description of the embodiments, those skilled in the art will clearly understand that the facilities of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Embodiments of the present invention can be implemented using existing processors, or by dedicated processors used for this or other purposes for suitable systems, or by hardwired systems. Embodiments of the present invention also include non-transitory computer-readable storage media, comprising machine-readable media for carrying or having machine-executable instructions or data structures stored thereon; such machine-readable media can be any available medium accessible by a general-purpose or special-purpose computer or other machine with a processor. For example, such machine-readable media can include RAM, ROM, EPROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store the required program code in the form of machine-executable instructions or data structures and is accessible by a general-purpose or special-purpose computer or other machine with a processor. When information is transmitted or provided to a machine via a network or other communication connection (hardwired, wireless, or a combination of hardwired and wireless), that connection is also considered a machine-readable medium.

[0211] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A multi-agent resource regulation method for airport terminals based on graph attention networks, characterized in that, The method includes the following steps: Constructing a terminal building process network model includes: obtaining the terminal building process topology, constructing a directed graph structure of resource nodes, and performing passenger generation and behavior, process advancement, and queuing calculations. A Markov game model is constructed to formalize the dynamic regulation problem of multiple resource nodes into a multi-agent reinforcement learning framework. The state space, observation space, action space, reward function and policy objective are determined. Under the multi-agent reinforcement learning framework, each agent corresponds to a resource node and performs resource regulation and configuration based on local and upstream and downstream information to minimize the average waiting time of passengers at each resource node and achieve the balance of terminal resource allocation. A multi-agent network structure perception mechanism for the terminal building is determined. This mechanism integrates two mechanisms: adaptive feature selection and hierarchical attention perception. It can dynamically extract and aggregate key information based on the state of the resource nodes themselves and their hierarchy and adjacency relationships in the network, thereby obtaining a state feature representation that takes into account both local features and global structure. Based on the obtained state characteristics of resource nodes, multi-agent intelligent decision generation is performed to obtain dynamic control decisions for terminal resource nodes.

2. The terminal multi-agent resource regulation method based on graph attention network according to claim 1, characterized in that, Adaptive feature selection mechanisms include: Define the original feature vector based on the node type; A feature selection attention mechanism is introduced, which automatically adjusts the importance of different features by calculating the correlation between node features and node type embeddings. The feature selection attention mechanism is expressed as follows: ； in, Represents resource nodes Feature selection attention weight vector; For resource nodes The original feature vector; For resource nodes Type embedding; Let be the learnable feature transformation matrix corresponding to the tanh activation function. This is the attention score parameter matrix before normalization using the softmax function; This represents a vector concatenation operation; Represents a nonlinear activation function; Represents the normalization function; The differentiated feature representation of resource nodes is obtained through weighted operations: ； in, Represents resource nodes Differentiated weighted feature representation vector; To achieve feature alignment between nodes of different types, a type-aware projection mechanism is further introduced to map differentiated features to a unified embedding space: ； in, Represents resource nodes The unified embedding representation vector obtained after type-aware projection; This represents a type-aware projection function; This represents a multi-layered sensing mechanism; This is the set of parameters for the MLP.

3. The terminal multi-agent resource regulation method based on graph attention network according to claim 2, characterized in that, Layered attention perception mechanisms include: A unified embedding representation vector for each resource node Mapping to hidden representation: ； in, Represents resource nodes Hidden representation of nodes in the hidden space; is a learnable linear mapping parameter matrix; To describe the hierarchical dependencies between nodes, a gating factor is introduced: ； in, Represents resource nodes With resource nodes The hierarchical perception gating coefficients between them; For the Sigmoid function; A vector of learnable parameters representing the gating factor; and Representing resource nodes respectively and resource nodes The hidden representation; and Representing resource nodes respectively With resource nodes The embedding vector of the corresponding level; Based on this, we define attention weights: ； in, This indicates the neighboring resource nodes at the current time step. For resource nodes Hierarchical perception attention weights; This represents the learnable parameter vector in the attention mechanism; These represent the hidden representations of resource node k, respectively. Represents resource nodes The set of neighboring nodes in the terminal process network; resource nodes The result of hierarchical sensing feature aggregation is as follows: ； in, Represents resource nodes Hierarchical perception feature representation after fusing multi-level structural information of neighboring nodes; This indicates a modified linear unit activation function; Finally, the state feature representation of the resource nodes in the fused hierarchical structure is obtained. : 。 4. The terminal multi-agent resource regulation method based on graph attention network according to claim 1, characterized in that, The reward function is: ； in, Representative resource node At time step The total reward value obtained, Represents local rewards, Represents collaborative rewards, Represents fair rewards; in, ； ； ； ； in, For resource nodes At time step Average waiting time at any given moment; This represents the total number of system resource nodes. This is the current resource node; This represents the global average latency. Indicates the current time step; This represents the number of discrete time steps in the statistics.

5. The terminal multi-agent resource regulation method based on graph attention network according to claim 1, characterized in that: The state space is composed of all the static and dynamic attributes of the resource nodes, and is represented as follows: ； in, This represents the set of resource nodes within the terminal building; Identifier for resource node type; For resource node level identification; Real-time queue length for resource nodes; This represents the average waiting time for the node. For the current resource node Number of open channels; For the current resource node The actual land area occupied; For the current resource node At time step The average floor space per passenger at any given time; The observation space is used to form an observation vector for each resource node according to its own node type, and further integrates the state information from neighboring resource nodes to reflect the local operating state of the node at the current moment and its upstream and downstream relationships. The action space is for adjusting the number of open channels of the current resource node, and is represented by a finite discrete set {-1,0,1} to reduce, maintain or increase one service channel; The strategy aims to have each agent generate actions based on local observations, with the goal of maximizing the expected discounted reward.

6. The terminal multi-agent resource regulation method based on graph attention network according to claim 1, characterized in that, Based on the obtained state characteristics of resource nodes, a multi-agent intelligent decision-making method is generated to obtain a dynamic control decision for terminal resource nodes: State information reception and fusion: At each decision time step, the agents of each functional node in the terminal receive state information from the environment; Policy network decision generation: Input state features into the policy network agent, extract key decision features through deep neural structure and output the optimal control action; in the terminal resource control scenario, the agent outputs the action of adjusting the number of channels or counters that can be opened at the current node; Action execution and environmental feedback: The agent applies the generated control actions to the terminal simulation model, and the operating state of the corresponding node changes accordingly; the environment calculates real-time feedback information based on the new operating state, including the reward value of the node and the global system, and outputs the observation state of the next time step. The real-time feedback information reflects the impact of the current decision on passenger throughput efficiency and the overall system balance. Experience trajectory recording and strategy optimization preparation: After completing a decision and interaction with the environment, the system will record the state, action, reward and next state information of each agent to form a complete experience trajectory sequence; By repeating the above steps, the algorithm continuously updates and optimizes the strategy parameters in multiple rounds of iteration, eventually converging to the optimal control scheme. This enables the coordination and dynamic optimization of multiple resource nodes in the terminal, and obtains dynamic control decisions for the terminal resource nodes.

7. The terminal multi-agent resource regulation method based on graph attention network according to claim 1, characterized in that, When performing process advancement and queue calculation... The system advances the simulation in fixed time steps, dynamically updating passenger status and node actions; passengers Waiting time Defined as: ； Average wait time of resource nodes for: ； in, Indicates the passenger queue Passengers ahead, please gather. For passengers Service hours, For resource nodes The number of channels; Represents resource nodes The first One service channel This indicates the resource nodes within the current statistical time window. The A group of passengers gathered in the waiting area. Indicates channel The number of passengers included in the statistics.

8. A terminal multi-agent resource control system based on graph attention networks, characterized in that, include: The terminal process network model construction module is used to obtain the terminal process topology and construct a directed graph structure of resource nodes to perform passenger generation and behavior, process advancement and queuing calculation. The Markov game model construction module is used to formalize the dynamic regulation problem of multiple resource nodes into a multi-agent reinforcement learning framework, which determines the state space, observation space, action space, reward function and policy objective. Under the multi-agent reinforcement learning framework, each agent corresponds to a resource node and performs resource regulation and configuration based on local and upstream and downstream information to minimize the average waiting time of passengers at each resource node and achieve a balance in terminal resource allocation. The structure perception mechanism determination module is used to determine the structure perception mechanism of the terminal multi-agent network. The mechanism integrates two mechanisms: adaptive feature selection and hierarchical attention perception. It can dynamically extract and aggregate key information based on the state of the resource node itself and its hierarchy and adjacency relationship in the network, thereby obtaining a state feature representation that takes into account both local features and global structure. The dynamic control decision acquisition module is used to generate multi-agent intelligent decisions based on the obtained state characteristics of resource nodes, and obtain dynamic control decisions for terminal resource nodes.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the terminal multi-agent resource regulation method based on graph attention networks as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the terminal multi-agent resource regulation method based on graph attention network as described in any one of claims 1 to 7.

Images